No. 01 · 1 page
BetterEnv World
Before post-training, give the agent environment a health check. A quality-diagnosis protocol that separates no signal, fake signal, and hackable rewards — plus a multi-stage multi-agent debug pipeline.
Research Notes
Papers from RsiingSI
the opensource org that scaling the rsi of harness\model\environment.
Short, citable notes on agent environments, benchmarks, and what recursive self-improvement actually requires — typeset as one- or two-page conference-style PDFs. Every claim is a position we are willing to publish before the experiments are finished. Read the manifesto ↓
Environments, Benchmarks & Infra
05
No. 02 · 2 pages
HAVEN: Harness Attribution
On a domain benchmark the model is fixed, so what is really being compared is the harness. HAVEN separates harness mechanism quality from the tool dividend, under three controlled conditions.
No. 03 · 1 page
The QA-GUI Environment Equivalence
Every QA pair is a flattened GUI environment: the question is the instruction, the context is the interface state, the answer is the reward. Unfold it and static health QA becomes executable, cross-device RL data.
No. 04 · 1 page
rsibio-paper2env
bioRxiv, PubMed / MEDLINE and Europe PMC hold the raw material. Filter through a multi-agent pipeline, check, difficulty-augment in the spirit of evolve-instruct, then build harbor-format environment data.
No. 05 · 1 page
RSI for AGI Infra
Recursive self-improvement has to combine digital and physical agents — if a model can optimize training infrastructure, why should the machine room not be maintained by agents too? A mixed, long-horizon benchmark to measure the distance.
Discovery & Meta-science
02
No. 06 · 1 page
ConceptSpace-Explorer
Summarizing and retrieving over existing papers cannot escape literature gravity. Instead of walking the literature, search concept space for structural blanks that should be connected but are not.
No. 07 · 1 page
Fast-Influence Research
Publish the problem and the first idea on xhs or X, let the paper come later, and treat being scooped as proof of taste. Idea citation as a mode of influence in the post-agent era.
Manifesto
Fast-influence research · Idea citationAfter reading the “slow science” proposal, I personally believe the post-agent era will bring a new mode of arXiv slot-claiming and academic-influence building — which I call “fast-influence research”, or idea citation research:
- Post the problem you have observed and your initial solution idea on xhs or Twitter/X.
- Let the experiments and the complete paper be advanced step by step later.
- Even if someone sees your idea and rushes out a paper with a similar core…
- …it does not really matter — you should be glad instead.
- Because it proves that you have foresight, and the taste to find problems and solution directions.
- And if one insists on citations, this can in fact form an idea-citation mode; I am certain this kind of community atmosphere will emerge.
I think this is the most important capability in the agent era: the foresight to raise valuable problems and initial directions.
The ability to implement will gradually be aligned onto agents for everyone — ultimately a resource-level contest of tokens, compute, and staying up late. But on reflection this is a deeper waste: people hold similar ideas, and someone else gets the publication out simply because they have a little more resource.
This is precisely the research mode of RsiingSI. Because we want to solve the real life-science problems of humanity, if we can throw the problems and the initial ideas out first, then anyone who sees them and is interested only needs to build them — that is one small step for humanity, and only this can truly accelerate the overall evolution of human society at maximum speed.
just for your imformation dm if you interest it